[Bug target/97832] AoSoA complex caxpy-like loops: AVX2+FMA -Ofast 7 times slower than -O3

["already5chosen at yahoo dot com" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=97832

--- Comment #3 from Michael_S <already5chosen at yahoo dot com> ---
(In reply to Richard Biener from comment #2)
> It's again reassociation making a mess out of the natural SLP opportunity
> (and thus SLP discovery fails miserably).
> 
> One idea worth playing with would be to change reassociation to rank
> references
> from the same load group (as later vectorization would discover) the same.
> 
> That said, further analysis and maybe a smaller testcase to look at is useful
> here.  There is, after all, the opportunity to turn "bad" association at the
> source level to good for vectorization when -ffast-math is enabled as well.

It turned out, much simpler kernel suffers from the same problem.

void foo1x1(double* restrict y, const double* restrict x, int clen)
{
  int xi = clen & 2;
  double f_re = x[0+xi+0];
  double f_im = x[4+xi+0];
  int clen2 = (clen+xi) * 2;
  #pragma GCC unroll 0
  for (int c = 0; c < clen2; c += 8) {
    // y[c] = y[c] - x[c]*conj(f);
    #pragma GCC unroll 4
    for (int k = 0; k < 4; ++k) {
      double x_re = x[c+0+k];
      double x_im = x[c+4+k];
      double y_re = y[c+0+k];
      double y_im = y[c+4+k];
      y_re = y_re - x_re * f_re - x_im * f_im;;
      y_im = y_im + x_re * f_im - x_im * f_re;
      y[c+0+k] = y_re;
      y[c+4+k] = y_im;
    }
  }
}

May be, it's possible to simplify further, but probably not by much.